layers.py

# coding=utf-8
# Copyright 2020 The Google Research Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# pylint: skip-file
"""Common layers for defining score networks.
"""
import math
import string
from functools import partial
import torch.nn as nn
import torch
import torch.nn.functional as F
import numpy as np
from .normalization import ConditionalInstanceNorm2dPlus


def get_act(config):
  """Get activation functions from the config file."""

  if config.model.nonlinearity.lower() == 'elu':
    return nn.ELU()
  elif config.model.nonlinearity.lower() == 'relu':
    return nn.ReLU()
  elif config.model.nonlinearity.lower() == 'lrelu':
    return nn.LeakyReLU(negative_slope=0.2)
  elif config.model.nonlinearity.lower() == 'swish':
    return nn.SiLU()
  else:
    raise NotImplementedError('activation function does not exist!')


def ncsn_conv1x1(in_planes, out_planes, stride=1, bias=True, dilation=1, init_scale=1., padding=0):
  """1x1 convolution. Same as NCSNv1/v2."""
  conv = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=bias, dilation=dilation,
                   padding=padding)
  init_scale = 1e-10 if init_scale == 0 else init_scale
  conv.weight.data *= init_scale
  conv.bias.data *= init_scale
  return conv


def variance_scaling(scale, mode, distribution,
                     in_axis=1, out_axis=0,
                     dtype=torch.float32,
                     device='cpu'):
  """Ported from JAX. """

  def _compute_fans(shape, in_axis=1, out_axis=0):
    receptive_field_size = np.prod(shape) / shape[in_axis] / shape[out_axis]
    fan_in = shape[in_axis] * receptive_field_size
    fan_out = shape[out_axis] * receptive_field_size
    return fan_in, fan_out

  def init(shape, dtype=dtype, device=device):
    fan_in, fan_out = _compute_fans(shape, in_axis, out_axis)
    if mode == "fan_in":
      denominator = fan_in
    elif mode == "fan_out":
      denominator = fan_out
    elif mode == "fan_avg":
      denominator = (fan_in + fan_out) / 2
    else:
      raise ValueError(
        "invalid mode for variance scaling initializer: {}".format(mode))
    variance = scale / denominator
    if distribution == "normal":
      return torch.randn(*shape, dtype=dtype, device=device) * np.sqrt(variance)
    elif distribution == "uniform":
      return (torch.rand(*shape, dtype=dtype, device=device) * 2. - 1.) * np.sqrt(3 * variance)
    else:
      raise ValueError("invalid distribution for variance scaling initializer")

  return init


def default_init(scale=1.):
  """The same initialization used in DDPM."""
  scale = 1e-10 if scale == 0 else scale
  return variance_scaling(scale, 'fan_avg', 'uniform')


class Dense(nn.Module):
  """Linear layer with `default_init`."""
  def __init__(self):
    super().__init__()


def ddpm_conv1x1(in_planes, out_planes, stride=1, bias=True, init_scale=1., padding=0):
  """1x1 convolution with DDPM initialization."""
  conv = nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=padding, bias=bias)
  conv.weight.data = default_init(init_scale)(conv.weight.data.shape)
  nn.init.zeros_(conv.bias)
  return conv


def ncsn_conv3x3(in_planes, out_planes, stride=1, bias=True, dilation=1, init_scale=1., padding=1):
  """3x3 convolution with PyTorch initialization. Same as NCSNv1/NCSNv2."""
  init_scale = 1e-10 if init_scale == 0 else init_scale
  conv = nn.Conv2d(in_planes, out_planes, stride=stride, bias=bias,
                   dilation=dilation, padding=padding, kernel_size=3)
  conv.weight.data *= init_scale
  conv.bias.data *= init_scale
  return conv


def ddpm_conv3x3(in_planes, out_planes, stride=1, bias=True, dilation=1, init_scale=1., padding=1):
  """3x3 convolution with DDPM initialization."""
  conv = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=padding,
                   dilation=dilation, bias=bias)
  conv.weight.data = default_init(init_scale)(conv.weight.data.shape)
  nn.init.zeros_(conv.bias)
  return conv

  ###########################################################################
  # Functions below are ported over from the NCSNv1/NCSNv2 codebase:
  # https://github.com/ermongroup/ncsn
  # https://github.com/ermongroup/ncsnv2
  ###########################################################################


class CRPBlock(nn.Module):
  def __init__(self, features, n_stages, act=nn.ReLU(), maxpool=True):
    super().__init__()
    self.convs = nn.ModuleList()
    for i in range(n_stages):
      self.convs.append(ncsn_conv3x3(features, features, stride=1, bias=False))
    self.n_stages = n_stages
    if maxpool:
      self.pool = nn.MaxPool2d(kernel_size=5, stride=1, padding=2)
    else:
      self.pool = nn.AvgPool2d(kernel_size=5, stride=1, padding=2)

    self.act = act

  def forward(self, x):
    x = self.act(x)
    path = x
    for i in range(self.n_stages):
      path = self.pool(path)
      path = self.convs[i](path)
      x = path + x
    return x


class CondCRPBlock(nn.Module):
  def __init__(self, features, n_stages, num_classes, normalizer, act=nn.ReLU()):
    super().__init__()
    self.convs = nn.ModuleList()
    self.norms = nn.ModuleList()
    self.normalizer = normalizer
    for i in range(n_stages):
      self.norms.append(normalizer(features, num_classes, bias=True))
      self.convs.append(ncsn_conv3x3(features, features, stride=1, bias=False))

    self.n_stages = n_stages
    self.pool = nn.AvgPool2d(kernel_size=5, stride=1, padding=2)
    self.act = act

  def forward(self, x, y):
    x = self.act(x)
    path = x
    for i in range(self.n_stages):
      path = self.norms[i](path, y)
      path = self.pool(path)
      path = self.convs[i](path)

      x = path + x
    return x


class RCUBlock(nn.Module):
  def __init__(self, features, n_blocks, n_stages, act=nn.ReLU()):
    super().__init__()

    for i in range(n_blocks):
      for j in range(n_stages):
        setattr(self, '{}_{}_conv'.format(i + 1, j + 1), ncsn_conv3x3(features, features, stride=1, bias=False))

    self.stride = 1
    self.n_blocks = n_blocks
    self.n_stages = n_stages
    self.act = act

  def forward(self, x):
    for i in range(self.n_blocks):
      residual = x
      for j in range(self.n_stages):
        x = self.act(x)
        x = getattr(self, '{}_{}_conv'.format(i + 1, j + 1))(x)

      x += residual
    return x


class CondRCUBlock(nn.Module):
  def __init__(self, features, n_blocks, n_stages, num_classes, normalizer, act=nn.ReLU()):
    super().__init__()

    for i in range(n_blocks):
      for j in range(n_stages):
        setattr(self, '{}_{}_norm'.format(i + 1, j + 1), normalizer(features, num_classes, bias=True))
        setattr(self, '{}_{}_conv'.format(i + 1, j + 1), ncsn_conv3x3(features, features, stride=1, bias=False))

    self.stride = 1
    self.n_blocks = n_blocks
    self.n_stages = n_stages
    self.act = act
    self.normalizer = normalizer

  def forward(self, x, y):
    for i in range(self.n_blocks):
      residual = x
      for j in range(self.n_stages):
        x = getattr(self, '{}_{}_norm'.format(i + 1, j + 1))(x, y)
        x = self.act(x)
        x = getattr(self, '{}_{}_conv'.format(i + 1, j + 1))(x)

      x += residual
    return x


class MSFBlock(nn.Module):
  def __init__(self, in_planes, features):
    super().__init__()
    assert isinstance(in_planes, list) or isinstance(in_planes, tuple)
    self.convs = nn.ModuleList()
    self.features = features

    for i in range(len(in_planes)):
      self.convs.append(ncsn_conv3x3(in_planes[i], features, stride=1, bias=True))

  def forward(self, xs, shape):
    sums = torch.zeros(xs[0].shape[0], self.features, *shape, device=xs[0].device)
    for i in range(len(self.convs)):
      h = self.convs[i](xs[i])
      h = F.interpolate(h, size=shape, mode='bilinear', align_corners=True)
      sums += h
    return sums


class CondMSFBlock(nn.Module):
  def __init__(self, in_planes, features, num_classes, normalizer):
    super().__init__()
    assert isinstance(in_planes, list) or isinstance(in_planes, tuple)

    self.convs = nn.ModuleList()
    self.norms = nn.ModuleList()
    self.features = features
    self.normalizer = normalizer

    for i in range(len(in_planes)):
      self.convs.append(ncsn_conv3x3(in_planes[i], features, stride=1, bias=True))
      self.norms.append(normalizer(in_planes[i], num_classes, bias=True))

  def forward(self, xs, y, shape):
    sums = torch.zeros(xs[0].shape[0], self.features, *shape, device=xs[0].device)
    for i in range(len(self.convs)):
      h = self.norms[i](xs[i], y)
      h = self.convs[i](h)
      h = F.interpolate(h, size=shape, mode='bilinear', align_corners=True)
      sums += h
    return sums


class RefineBlock(nn.Module):
  def __init__(self, in_planes, features, act=nn.ReLU(), start=False, end=False, maxpool=True):
    super().__init__()

    assert isinstance(in_planes, tuple) or isinstance(in_planes, list)
    self.n_blocks = n_blocks = len(in_planes)

    self.adapt_convs = nn.ModuleList()
    for i in range(n_blocks):
      self.adapt_convs.append(RCUBlock(in_planes[i], 2, 2, act))

    self.output_convs = RCUBlock(features, 3 if end else 1, 2, act)

    if not start:
      self.msf = MSFBlock(in_planes, features)

    self.crp = CRPBlock(features, 2, act, maxpool=maxpool)

  def forward(self, xs, output_shape):
    assert isinstance(xs, tuple) or isinstance(xs, list)
    hs = []
    for i in range(len(xs)):
      h = self.adapt_convs[i](xs[i])
      hs.append(h)

    if self.n_blocks > 1:
      h = self.msf(hs, output_shape)
    else:
      h = hs[0]

    h = self.crp(h)
    h = self.output_convs(h)

    return h


class CondRefineBlock(nn.Module):
  def __init__(self, in_planes, features, num_classes, normalizer, act=nn.ReLU(), start=False, end=False):
    super().__init__()

    assert isinstance(in_planes, tuple) or isinstance(in_planes, list)
    self.n_blocks = n_blocks = len(in_planes)

    self.adapt_convs = nn.ModuleList()
    for i in range(n_blocks):
      self.adapt_convs.append(
        CondRCUBlock(in_planes[i], 2, 2, num_classes, normalizer, act)
      )

    self.output_convs = CondRCUBlock(features, 3 if end else 1, 2, num_classes, normalizer, act)

    if not start:
      self.msf = CondMSFBlock(in_planes, features, num_classes, normalizer)

    self.crp = CondCRPBlock(features, 2, num_classes, normalizer, act)

  def forward(self, xs, y, output_shape):
    assert isinstance(xs, tuple) or isinstance(xs, list)
    hs = []
    for i in range(len(xs)):
      h = self.adapt_convs[i](xs[i], y)
      hs.append(h)

    if self.n_blocks > 1:
      h = self.msf(hs, y, output_shape)
    else:
      h = hs[0]

    h = self.crp(h, y)
    h = self.output_convs(h, y)

    return h


class ConvMeanPool(nn.Module):
  def __init__(self, input_dim, output_dim, kernel_size=3, biases=True, adjust_padding=False):
    super().__init__()
    if not adjust_padding:
      conv = nn.Conv2d(input_dim, output_dim, kernel_size, stride=1, padding=kernel_size // 2, bias=biases)
      self.conv = conv
    else:
      conv = nn.Conv2d(input_dim, output_dim, kernel_size, stride=1, padding=kernel_size // 2, bias=biases)

      self.conv = nn.Sequential(
        nn.ZeroPad2d((1, 0, 1, 0)),
        conv
      )

  def forward(self, inputs):
    output = self.conv(inputs)
    output = sum([output[:, :, ::2, ::2], output[:, :, 1::2, ::2],
                  output[:, :, ::2, 1::2], output[:, :, 1::2, 1::2]]) / 4.
    return output


class MeanPoolConv(nn.Module):
  def __init__(self, input_dim, output_dim, kernel_size=3, biases=True):
    super().__init__()
    self.conv = nn.Conv2d(input_dim, output_dim, kernel_size, stride=1, padding=kernel_size // 2, bias=biases)

  def forward(self, inputs):
    output = inputs
    output = sum([output[:, :, ::2, ::2], output[:, :, 1::2, ::2],
                  output[:, :, ::2, 1::2], output[:, :, 1::2, 1::2]]) / 4.
    return self.conv(output)


class UpsampleConv(nn.Module):
  def __init__(self, input_dim, output_dim, kernel_size=3, biases=True):
    super().__init__()
    self.conv = nn.Conv2d(input_dim, output_dim, kernel_size, stride=1, padding=kernel_size // 2, bias=biases)
    self.pixelshuffle = nn.PixelShuffle(upscale_factor=2)

  def forward(self, inputs):
    output = inputs
    output = torch.cat([output, output, output, output], dim=1)
    output = self.pixelshuffle(output)
    return self.conv(output)


class ConditionalResidualBlock(nn.Module):
  def __init__(self, input_dim, output_dim, num_classes, resample=1, act=nn.ELU(),
               normalization=ConditionalInstanceNorm2dPlus, adjust_padding=False, dilation=None):
    super().__init__()
    self.non_linearity = act
    self.input_dim = input_dim
    self.output_dim = output_dim
    self.resample = resample
    self.normalization = normalization
    if resample == 'down':
      if dilation > 1:
        self.conv1 = ncsn_conv3x3(input_dim, input_dim, dilation=dilation)
        self.normalize2 = normalization(input_dim, num_classes)
        self.conv2 = ncsn_conv3x3(input_dim, output_dim, dilation=dilation)
        conv_shortcut = partial(ncsn_conv3x3, dilation=dilation)
      else:
        self.conv1 = ncsn_conv3x3(input_dim, input_dim)
        self.normalize2 = normalization(input_dim, num_classes)
        self.conv2 = ConvMeanPool(input_dim, output_dim, 3, adjust_padding=adjust_padding)
        conv_shortcut = partial(ConvMeanPool, kernel_size=1, adjust_padding=adjust_padding)

    elif resample is None:
      if dilation > 1:
        conv_shortcut = partial(ncsn_conv3x3, dilation=dilation)
        self.conv1 = ncsn_conv3x3(input_dim, output_dim, dilation=dilation)
        self.normalize2 = normalization(output_dim, num_classes)
        self.conv2 = ncsn_conv3x3(output_dim, output_dim, dilation=dilation)
      else:
        conv_shortcut = nn.Conv2d
        self.conv1 = ncsn_conv3x3(input_dim, output_dim)
        self.normalize2 = normalization(output_dim, num_classes)
        self.conv2 = ncsn_conv3x3(output_dim, output_dim)
    else:
      raise Exception('invalid resample value')

    if output_dim != input_dim or resample is not None:
      self.shortcut = conv_shortcut(input_dim, output_dim)

    self.normalize1 = normalization(input_dim, num_classes)

  def forward(self, x, y):
    output = self.normalize1(x, y)
    output = self.non_linearity(output)
    output = self.conv1(output)
    output = self.normalize2(output, y)
    output = self.non_linearity(output)
    output = self.conv2(output)

    if self.output_dim == self.input_dim and self.resample is None:
      shortcut = x
    else:
      shortcut = self.shortcut(x)

    return shortcut + output


class ResidualBlock(nn.Module):
  def __init__(self, input_dim, output_dim, resample=None, act=nn.ELU(),
               normalization=nn.InstanceNorm2d, adjust_padding=False, dilation=1):
    super().__init__()
    self.non_linearity = act
    self.input_dim = input_dim
    self.output_dim = output_dim
    self.resample = resample
    self.normalization = normalization
    if resample == 'down':
      if dilation > 1:
        self.conv1 = ncsn_conv3x3(input_dim, input_dim, dilation=dilation)
        self.normalize2 = normalization(input_dim)
        self.conv2 = ncsn_conv3x3(input_dim, output_dim, dilation=dilation)
        conv_shortcut = partial(ncsn_conv3x3, dilation=dilation)
      else:
        self.conv1 = ncsn_conv3x3(input_dim, input_dim)
        self.normalize2 = normalization(input_dim)
        self.conv2 = ConvMeanPool(input_dim, output_dim, 3, adjust_padding=adjust_padding)
        conv_shortcut = partial(ConvMeanPool, kernel_size=1, adjust_padding=adjust_padding)

    elif resample is None:
      if dilation > 1:
        conv_shortcut = partial(ncsn_conv3x3, dilation=dilation)
        self.conv1 = ncsn_conv3x3(input_dim, output_dim, dilation=dilation)
        self.normalize2 = normalization(output_dim)
        self.conv2 = ncsn_conv3x3(output_dim, output_dim, dilation=dilation)
      else:
        # conv_shortcut = nn.Conv2d ### Something wierd here.
        conv_shortcut = partial(ncsn_conv1x1)
        self.conv1 = ncsn_conv3x3(input_dim, output_dim)
        self.normalize2 = normalization(output_dim)
        self.conv2 = ncsn_conv3x3(output_dim, output_dim)
    else:
      raise Exception('invalid resample value')

    if output_dim != input_dim or resample is not None:
      self.shortcut = conv_shortcut(input_dim, output_dim)

    self.normalize1 = normalization(input_dim)

  def forward(self, x):
    output = self.normalize1(x)
    output = self.non_linearity(output)
    output = self.conv1(output)
    output = self.normalize2(output)
    output = self.non_linearity(output)
    output = self.conv2(output)

    if self.output_dim == self.input_dim and self.resample is None:
      shortcut = x
    else:
      shortcut = self.shortcut(x)

    return shortcut + output


###########################################################################
# Functions below are ported over from the DDPM codebase:
#  https://github.com/hojonathanho/diffusion/blob/master/diffusion_tf/nn.py
###########################################################################

def get_timestep_embedding(timesteps, embedding_dim, max_positions=10000):
  assert len(timesteps.shape) == 1  # and timesteps.dtype == tf.int32
  half_dim = embedding_dim // 2
  # magic number 10000 is from transformers
  emb = math.log(max_positions) / (half_dim - 1)
  # emb = math.log(2.) / (half_dim - 1)
  emb = torch.exp(torch.arange(half_dim, dtype=torch.float32, device=timesteps.device) * -emb)
  # emb = tf.range(num_embeddings, dtype=jnp.float32)[:, None] * emb[None, :]
  # emb = tf.cast(timesteps, dtype=jnp.float32)[:, None] * emb[None, :]
  emb = timesteps.float()[:, None] * emb[None, :]
  emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
  if embedding_dim % 2 == 1:  # zero pad
    emb = F.pad(emb, (0, 1), mode='constant')
  assert emb.shape == (timesteps.shape[0], embedding_dim)
  return emb


def _einsum(a, b, c, x, y):
  einsum_str = '{},{}->{}'.format(''.join(a), ''.join(b), ''.join(c))
  return torch.einsum(einsum_str, x, y)


def contract_inner(x, y):
  """tensordot(x, y, 1)."""
  x_chars = list(string.ascii_lowercase[:len(x.shape)])
  y_chars = list(string.ascii_lowercase[len(x.shape):len(y.shape) + len(x.shape)])
  y_chars[0] = x_chars[-1]  # first axis of y and last of x get summed
  out_chars = x_chars[:-1] + y_chars[1:]
  return _einsum(x_chars, y_chars, out_chars, x, y)


class NIN(nn.Module):
  def __init__(self, in_dim, num_units, init_scale=0.1):
    super().__init__()
    self.W = nn.Parameter(default_init(scale=init_scale)((in_dim, num_units)), requires_grad=True)
    self.b = nn.Parameter(torch.zeros(num_units), requires_grad=True)

  def forward(self, x):
    x = x.permute(0, 2, 3, 1)
    y = contract_inner(x, self.W) + self.b
    return y.permute(0, 3, 1, 2)


class AttnBlock(nn.Module):
  """Channel-wise self-attention block."""
  def __init__(self, channels):
    super().__init__()
    self.GroupNorm_0 = nn.GroupNorm(num_groups=32, num_channels=channels, eps=1e-6)
    self.NIN_0 = NIN(channels, channels)
    self.NIN_1 = NIN(channels, channels)
    self.NIN_2 = NIN(channels, channels)
    self.NIN_3 = NIN(channels, channels, init_scale=0.)

  def forward(self, x):
    B, C, H, W = x.shape
    h = self.GroupNorm_0(x)
    q = self.NIN_0(h)
    k = self.NIN_1(h)
    v = self.NIN_2(h)

    w = torch.einsum('bchw,bcij->bhwij', q, k) * (int(C) ** (-0.5))
    w = torch.reshape(w, (B, H, W, H * W))
    w = F.softmax(w, dim=-1)
    w = torch.reshape(w, (B, H, W, H, W))
    h = torch.einsum('bhwij,bcij->bchw', w, v)
    h = self.NIN_3(h)
    return x + h


class Upsample(nn.Module):
  def __init__(self, channels, with_conv=False):
    super().__init__()
    if with_conv:
      self.Conv_0 = ddpm_conv3x3(channels, channels)
    self.with_conv = with_conv

  def forward(self, x):
    B, C, H, W = x.shape
    h = F.interpolate(x, (H * 2, W * 2), mode='nearest')
    if self.with_conv:
      h = self.Conv_0(h)
    return h


class Downsample(nn.Module):
  def __init__(self, channels, with_conv=False):
    super().__init__()
    if with_conv:
      self.Conv_0 = ddpm_conv3x3(channels, channels, stride=2, padding=0)
    self.with_conv = with_conv

  def forward(self, x):
    B, C, H, W = x.shape
    # Emulate 'SAME' padding
    if self.with_conv:
      x = F.pad(x, (0, 1, 0, 1))
      x = self.Conv_0(x)
    else:
      x = F.avg_pool2d(x, kernel_size=2, stride=2, padding=0)

    assert x.shape == (B, C, H // 2, W // 2)
    return x


class ResnetBlockDDPM(nn.Module):
  """The ResNet Blocks used in DDPM."""
  def __init__(self, act, in_ch, out_ch=None, temb_dim=None, conv_shortcut=False, dropout=0.1):
    super().__init__()
    if out_ch is None:
      out_ch = in_ch
    self.GroupNorm_0 = nn.GroupNorm(num_groups=32, num_channels=in_ch, eps=1e-6)
    self.act = act
    self.Conv_0 = ddpm_conv3x3(in_ch, out_ch)
    if temb_dim is not None:
      self.Dense_0 = nn.Linear(temb_dim, out_ch)
      self.Dense_0.weight.data = default_init()(self.Dense_0.weight.data.shape)
      nn.init.zeros_(self.Dense_0.bias)

    self.GroupNorm_1 = nn.GroupNorm(num_groups=32, num_channels=out_ch, eps=1e-6)
    self.Dropout_0 = nn.Dropout(dropout)
    self.Conv_1 = ddpm_conv3x3(out_ch, out_ch, init_scale=0.)
    if in_ch != out_ch:
      if conv_shortcut:
        self.Conv_2 = ddpm_conv3x3(in_ch, out_ch)
      else:
        self.NIN_0 = NIN(in_ch, out_ch)
    self.out_ch = out_ch
    self.in_ch = in_ch
    self.conv_shortcut = conv_shortcut

  def forward(self, x, temb=None):
    B, C, H, W = x.shape
    assert C == self.in_ch
    out_ch = self.out_ch if self.out_ch else self.in_ch
    h = self.act(self.GroupNorm_0(x))
    h = self.Conv_0(h)
    # Add bias to each feature map conditioned on the time embedding
    if temb is not None:
      h += self.Dense_0(self.act(temb))[:, :, None, None]
    h = self.act(self.GroupNorm_1(h))
    h = self.Dropout_0(h)
    h = self.Conv_1(h)
    if C != out_ch:
      if self.conv_shortcut:
        x = self.Conv_2(x)
      else:
        x = self.NIN_0(x)
    return x + h